A group of leading specialists from OpenAI, DeepMind, Meta and Salesforce joined the Recursive Superintelligence project, a startup that in a few months has attracted more than $500 million in investments at a valuation of $4 billion, writes xrust. The company is developing AI that can independently improve its models and learning methods.
- What happened: the startup gathered hundreds of millions and the best industry specialists
- Who is behind the project: a team of AI “star team” level
- Who invested: Google Ventures, Nvidia, AMD and others
- What is “recursive self-improvement” and why is it important
- Why this could change the industry
- What is already known about RSI technology
- Why researchers believe that the time has come
- But there are also skeptics: “a person is still needed”
- Why the project is of such interest in Silicon Valley
- When to expect the first results
- Background: who else is working on self-improving AI
What happened: the startup gathered hundreds of millions and the best industry specialists
According to The New York Times and Financial Times, Recursive Superintelligence (RSI), a young company founded in late 2025, has attracted more than $500 million in funding at a valuation of about $4 billion. The project includes world-class researchers who previously worked at OpenAI, Google DeepMind, Meta and Salesforce.
The startup sets itself an ambitious goal: to create an artificial intelligence system that can independently improve its own algorithms , reducing human participation in the development process.
Who is behind the project: a team of AI “star team” level
Recursive Superintelligence was founded by eight famous researchers, including:
- class=»notranslate»>__GTAG11__ Richard Socher — former chief scientist at Salesforce;
- Tim Rockteschel — professor at UCL, ex-DeepMind employee;
- Josh Tobin, Jeff Clune, Tim Shea — former OpenAI researchers;
- Yuandong Tian — AI specialist from Meta;
- Peter Norvig — legendary director of research at Google and co-author of the textbook Artificial Intelligence: A Modern Approach .
The team has less than 30 people, but has already attracted the attention of major investors.
Who invested: Google Ventures, Nvidia, AMD and others
The investment round was led by the fund GV (Google Ventures) , which was joined by Nvidia , AMD , Greycroft and other major players.
According to Tech Funding News, investor interest was so high that the round turned out to be oversubscribed — there were more people willing to invest than the company was willing to accept.
What is “recursive self-improvement” and why is it important
Recursive Superintelligence is working on a concept that in Silicon Valley is called recursive self‑improvement — recursive self-improvement.
The essence of the idea:
- The AI itself chooses the directions of research;
- generates hypotheses itself;
- conducts experiments himself;
- evaluates the results itself;
- itself improves its own models and architectures.
In fact, this is an attempt to create an artificial researcher who is able to work without constant human participation.
Why this could change the industry
If Recursive Superintelligence succeeds, it could:
- speed up the development of new AI models significantly;
- reduce the need for large teams of engineers;
- make AI development more autonomous;
- lead to the emergence of systems that create new learning methods faster than people do.
Experts call this “the next stage in the evolution of AI” — a transition from models that perform tasks to models that create new models .
What is already known about RSI technology
According to AIDirectory and ICO Optics, Recursive Superintelligence is working on a system that can:
- automatically select training data;
- run thousands of experiments in parallel;
- analyze the results and adjust hyperparameters;
- rewrite your own code;
- form new neural network architectures;
- work in the “open-endedness” mode. Open‑endedness is an approach in which AI works for months or years, constantly generating new ideas and improvements.
Why researchers believe that the time has come
Richard Socher, one of the founders, said:
“AI is code. And now AI can write code. All the ingredients for self-improvement already exist.”
According to Socher, modern models are already capable of:
- write complex programs;
- analyze your own errors;
- suggest improvements;
- learn from your own conclusions.
The next step is to combine these capabilities into a single system.
But there are also skeptics: “a person is still needed”
Some researchers believe that fully autonomous AI is not yet possible.
They point out that:
- AI is not able to formulate truly new scientific ideas;
- models are prone to errors and “hallucinations”;
- without a human, AI may choose the wrong direction of development;
- autonomous systems require serious monitoring and regulation.
However, even skeptics admit that automation of routine tasks in AI research is an inevitable trend.
Why the project is of such interest in Silicon Valley
Reasons:
- race for AGI (universal artificial intelligence) is enhanced;
- companies are looking for ways to speed up research;
- investors want to invest in the “next OpenAI”;
- recursive self-improvement is a long-standing dream of techno-optimists.
According to the NYT, Recursive Superintelligence is one of the most talked about projects of 2026.
When to expect the first results
The company plans a public launch in mid-May 2026 , where it will present the first prototypes and research tools.
Experts believe that in the next 1-2 years RSI may show:
- autonomous data selection systems;
- AI laboratories working without human intervention;
- models that themselves design next-generation architectures.
Background: who else is working on self-improving AI
According to DNYUZ, similar ideas are being worked on by:
- Anthropic — explores models that can independently correct behavior;
- OpenAI — develops systems that write and test code;
- AMI Labs (Yann LeCun) — creates “world models”;
- Ineffable Intelligence (David Silver) — working on AI that learns through self-exploration.
But Recursive Superintelligence is the only startup that relies specifically on full automation of AI research .
According to the news feed https://www.nytimes.com
Xrust Famous researchers have joined a $4 billion project to create self-improving artificial intelligence
